Only 10 years ago, barely any machine could reliably provide language or image recognition. Today, machines have learned to outperform humans on many tasks. In the past few months, we’ve seen progress in AI capabilities that has impressed even skeptics. A “golden decade,” one researcher called it. In 2023 and beyond, we will see more such systems (especially Generative AI systems like ChatGPT) complementing or replacing us human creators in many areas.
Artificial intelligence (AI). Even though I’ve worked extensively with this technology for 10 years, we are amid a significant leap forward in artificial intelligence. Just in the last few months, we’ve seen advances in AI capabilities that have impressed even skeptics. But first, let’s take a few more steps back and begin by looking at AI’s development in the past decade.
2012-2014 – The beginnings of image recognition, reading comprehension, and language understanding
Some researchers say that the year 2012 was a milestone for deep learning. It was the year when Google researchers built a large neural network of 16.000 processors with one billion connections to recognize pictures and videos of cats. This is an example of reinforcement learning which, besides supervised learning and probabilistic program induction, was one of the AI frameworks that were massively successful in the past decade. Recognizing cat pictures may seem like a minor achievement. However, at the time, machines just began to use deep learning for providing image recognition. In 2012, image recognition was still in its infancy, and in tests that compared AI to human performance, it was found that AI performed at about -40, which was still below human performance (here, set at the zero-base line). Besides image recognition, a decade ago, AI also still underperformed human beings on other tasks, including reading comprehension and language understanding. Even though in 2013, NELL (Never-Ending Language Learning), a semantic machine learning system, was invented, AI could still not reliably fulfill tasks regarding language processing. The ability of AI to recognize speech drastically improved with the invention of Alexa in 2014. Before that, Apple’s Siri allowed users to manage their phones via speech. Yet, AI’s ability to understand the language was still worse than the ability of human beings. In the following years, AI reached a better-than-human performance level in language understanding. This was due to the improvements in AI’s voice recognition, advances in language processing and related neural network language models, and information organization. Although AI systems still struggle to produce long and coherent texts, chatbots such as “ChatGPT” show the immense progress that has been made up to today.
2015 – 2017 AI began performing better than human beings
2015 marked the year when everyone was allowed to build meaningful AI models. After IBM’s signature artificial intelligence system “Watson” became famous in 2011 for beating Jeopardy! champion Ken Jennings, several open-source platforms for machine learning appeared on the market (such as Google’s open-sourced deep learning framework TensorFlow). It enabled companies and developers to work with the technology in new ways. Additionally, this year crucial progress was made regarding face and image recognition. For example, machines beat humans at the 6th edition of the ImageNet Large Scale Visual Recognition Challenge (ImageNet is a standardized collection of millions of photographs that trains and tests visual identification programs). Going forward, in 2016, deep reinforcement learning – a combination of neural networks and reinforcement learning – generated massive hype in the AI community when Google’s AlphaGo beat the world’s best Go player. Furthermore, in 2017, the use of self-supervised learning models in conjunction with deep neural networks was ramped up with the introduction of the Transformer. Today, these transformer models are the mainstream approach for Natural Language Processing (NLP), including applications such as machine translation and Google web search.
2018 – 2019 Data Security, Language Processing, and AI in Medicine
Due to the Cambridge Analytica scandal, 2018 was the year when the topic of data security reached its peak. In line with this, a McKinsey survey found that in 2018, the risk was one of the functions where most respondents said the value of AI was visible. Beyond this, when BERT was designed, language processing made a major leap in 2018. BERT is one example of a neural network language model that learns about using words, grammar, meaning, and basic facts in different contexts. By connecting word sequences simultaneously instead of stringing them together from left to right, models like BERT can generate abstracts almost indistinguishable from the human-generated text. These language processing models are crucial for supporting applications such as chatbots and have determined their immense progress over the past decade. Besides this, in 2019, researchers started creating an AI system that outperforms human radiologists in detecting lung cancer. This was achieved with a deep-learning algorithm that can interpret computed tomography (CT) scans to predict someone’s likelihood of having the disease.
2020 – 2021 Quick AI advancements due to the pandemic
During 2020, AI development was driven and accelerated due to the COVID-19 pandemic. Artificial Intelligence was greatly responsible for speeding up vaccine development which would have usually taken several decades. Instead, this process was significantly shortened because AI supported researchers in analyzing enormous amounts of data. The growth of AI is exemplified by global corporate investments of 68 billion USD, an increase of 40% from 2019 to 2020. Moreover, in 2021 alone, the number of patent filings related to AI innovations was 30 times larger than the number of filings in 2015, showing the rapid advancements in AI development. Last year the research community mainly focused on the application of AI to computers. This subfield teaches machines to comprehend images and other visual material to perform well at classifying images, recognizing objects, mapping the position and movement of human body joints, and detecting faces.
Today – AI has become indispensable in our lives.
Artificial Intelligence (AI) has undergone rapid development in the past decade. Only 10 years ago, barely any machine could reliably provide language or image recognition. Today, machines have learned to outperform humans on many tasks. For instance, AI systems can detect fraudulent charges before you know you’ve lost the card or check eligibility criteria when someone applies for a loan. AI recognizes patterns and evaluates options in our everyday lives. Which Instagram post I prefer and keep me on the social media platform, which price on Amazon makes me want to buy, and whether I wanted to leave my AirPods at home. In recent months, especially “generative AI” has exploded – systems that create new possibilities. We’ve seen progress in AI capabilities that has left even skeptics impressed.
A firework of Generative Artificial Intelligence
Within a few months of one another, three different image-generating AIs were made public: Dall-E, Midjourney, and Stable Diffusion. You enter a text, and the system generates an image within a few seconds. Ask for “an astronaut riding a horse on Mars,” and the AI gets started. One image of Midjourney was on the cover of the Economist in June 2022, and another won an award at the Colorado State Fair. A sign of what’s to come in 2023.
For me, the developments in text-generating AIs are even more exciting. The American startup OpenAI has triggered worldwide hype with its chatbot “ChatGPT.” This chatbot imitates the neural network of a human brain. The bot can have human-like conversations and generate convincing answers even to more complex technical questions and dialogues. This represents a massive milestone for AI systems because language processing has proven to be one of the most challenging tasks for AI in the past.
AI adoption in companies is on the rise.
This year’s State of AI report shows the accelerating progress also in the industry. In 2020, not a single drug was in clinical trials that had been developed using an AI-first approach. Today, there are 18. An AI system from BioNTech successfully identified numerous high-risk covid variants months before WHO’s tracking system detected them.
According to McKinsey, AI adoption has more than doubled since 2017. Robotic automation and computer vision are the AI capabilities most frequently deployed each year. Moreover, investment in AI has increased. Today 52% of respondents report that 5% of their digital budget is used for AI investments. In 2017, that proportion of respondents was only 40%.
However, AI adoption is still concentrated amongst the AI high performers. This implies that those companies that were leading with regard to AI, are still expanding their competitive advantage today. These AI leaders engage more in the “industrialization of AI” and linking their AI strategy to their core business practices. The reasons why these companies continue to outperform are more investments and higher spending compared to AI laggards which then attract more and better-performing (tech) talent.
What is to come for AI in the future?
In 2023, we will see more such systems complementing or replacing us human creators in many areas. Video generation is under development, and so is the creation of customized media. (“Siri, show me a 90-minute movie about a CEO who quits her career to making her fortune as an actress, Steven Spielberg style.”)
But although there has been tremendous progress and although more companies have adopted AI, a lot of research must only occur once we can produce fully general AI systems. In the next few years, we will see further progress regarding self-supervised learning models, continuous learning, and task generalization. For instance, future applications of neural network language models such as BERT could allow for human interaction across different languages and contexts. We will likely obtain higher quality data for languages used by smaller populations and that biases in AI systems will more easily be detected and removed. Therefore, we can also expect that AI chatbots such as “ChatGPT” will become more human-like and that they will be more widely applied in, for example schools and universities.
In the upcoming years, we will observe an extended use of facial recognition for control access and safety measures. China is one of the most prominent examples where the government has almost fully embedded facial recognition technology in its society. Nevertheless, it remains to be seen whether more countries will follow the Chinese example since this omnipresence of AI in societies will likely raise privacy, surveillance, and ethical concerns.
Certainly, Artificial Intelligence adoption and investment will continue to grow in the following years. In fact, as shown by the McKinsey survey, 63% of the respondents say that their AI investments will increase over the next years, creating benefits such as positive revenue effects and reduced costs. It is expected that by 2025, worldwide revenues of enterprise applications in the field of AI will rise to a level of 31 billion USD. However, AI growth needs to be transparent, and it needs to be controlled to deal with the privacy and ethical concerns. Therefore, governments and companies must create AI laws today that will guide the AI development in the future.